A Dashboard to Analysis and Synthesis of Dimensionality Reduction Methods in Remote Sensing

2025-04-30 0 0 153.86KB 7 页 10玖币

侵权投诉

A Dashboard to Analysis and Synthesis of

Dimensionality Reduction Methods in

Remote Sensing

Elkebir Sarhrouni #1, Ahmed Hammouch *2, Driss Aboutajdine#3

#LRIT, Faculty of Sciences, Mohamed V - Agdal University, Morocco,

A. Ibn Battouta. 4. B.P. : 1014 Rabat, Morocco

1 sarhrouni436@yahoo.fr

* LRGE, ENSET, Mohamed V - Souissi University, Morocco

A. FAR BP : 6207 - Rabat Instituts Rabat Morocco

Abstract—Hyperspectral images (HSI) classification is a high technical remote sensing software. The

purpose is to reproduce a thematic map . The HSI contains more than a hundred hyperspectral measures,

as bands (or simply images), of the concerned region. They are taken at neighbors frequencies.

Unfortunately, some bands are redundant features, others are noisily measured, and the high

dimensionality of features made classification accuracy poor. The problematic is how to find the good

bands to classify the regions items. Some methods use Mutual Information (MI) and thresholding, to

select relevant images, without processing redundancy. Others control and avoid redundancy. But they

process the dimensionality reduction, some times as selection, other times as wrapper methods without

any relationship . Here , we introduce a survey on all scheme used, and after critics and improvement, we

synthesize a dashboard, that helps user to analyze an hypothesize features selection and extraction

softwares.

Keyword-Feature Selection Software, Feature Extraction Software, Hyperspectral images

Classification, Remote Sensing.

I. INTRODUCTION

Due to the recent achievements in the remote sensing technologies, we are faced at large quantity of

information, organized at bidirectional measures of the same region, called bands, and taken at very closed

frequencies. The goal here is to determine patterns in order to classify the points and produce the thematic map

of the concerned region. This technology is called Hyperspectral Images (HSI), and it opens new applications

fields and renews the problematics posed in classification domain. Explicitly we are faced at reduction of

dimensionality problematic. the feature classification domain, the choice of data affects widely the results.So,

the bands don’t all contain the information; some ands are irrelevant like those affected by various atmospheric

effects,and decrease the classification accuracy. And there exist redundant ands to complicate the learning

system and product incorrect prediction [1].

Even the bands contain enough information about the scene they may can’t predict the classes correctly if the

dimension of space images, is so large that needs many cases to detect the relationship between the bands and

the scene (Hughes phenomenon) [5]. We can reduce the dimensionality of hyperspectral images by selecting

only the relevant bands (feature selection or subset selection methodology), or extracting, from the original

bands, new bands containing the maximal information about the classes, using any functions, logical or

numerical (feature extraction methodology) [4,6], or we can use an hybrid schema containing selection before

extraction.

An example of Hyperspectral image that largely served for academic search is AVIRIS 92AV3C (Airborne

Visible Infrared Imaging Spectrometer). [2]. It contains 220 images taken of the region "Indiana Pine" at

"north-western Indiana", USA [2]. The 220 called bands are taken between 0.4μm and 2.5μm. Each band has

145 lines and 145 columns. The ground truth map is also provided, but only 10366 pixels are labeled fro 1 to 16.

Each label indicates one from 16 classes. The zeros indicate pixels how are not classified yet, see Figure.1.

Elkebir Sarhrouni et.al / International Journal of Engineering and Technology (IJET)

ISSN : 0975-4024

Vol 5 No 3 Jun-Jul 2013

2678

Fig. 1. The Ground Truth map of AVIRIS 92AV3C and the 16 classes.

The hyperspectral image AVIRIS 92AV3C contains numbers (measures) between 955 and

9406. Each pixel of the ground truth map has a set of 220 numbers along the hyperspectral image. This

numbers (measures) represent the reflectance of the pixel in each band.

So reducing dimensionality means selecting only the dimensions caring a lot of information regarding the

classes.

In this paper we collect all techniques and algorithms related to dimensionality reduction applied to

Hyperspectral images, or that can be applied to HSI. This survey provides a work board for analysis and

synthesis of dimensionality reduction.

We start, in section two by citing the principle notions related at classification in the context of dimensionality

reduction. In section three, we develop different schema and strategies for features selection and extraction, and

we discuss their ability to be applied to HSI. The forth section provides our synthesis of a work board flow for

selection, extraction and classification of HSI remote sensing

II. DIMENSIONALITY REDUCTION

A. General Purpose

AN important question that often arises in the field of data mining is the problem of having too many attributes.

In other words, the measured attributes are not all likely to be necessary for accurate discrimination. Therefore

include some features the model classification can lead to a worse model than if they were removed. In addition

it is not clear if the features are (or are not) relevant. In this context, dimensionality reduction use data-adaptive

methods using a priori information available, to inform us which variables are clearly relevant or not to the task

of classification [3].

B. The Curse of Dimensionality

While, theoretically, having more features should give us more power to discriminate classes, the real world

gives us many reasons why this is not usually the case. This means that what may work well in a unidimensional

may not be extended to high dimensions. This is the case where the amount of data must be increased to

maintain a level of precision parameters.

Thus, in a task induction, when the number of attributes increases, the time required for an algorithm sometimes

grows exponentially. Therefore, when the set of attributes is large enough, the induction algorithms are simply

insurmountable. So that the induction methods suffer from the curse of dimensionality [5].

This problem is compounded by the fact that many attributes can be either irrelevant or redundant with other

features in predicting the class of an instance. In this context, these attributes serve no purpose except to

increase the s induction time.

C. Generic Process for Dimensionality Reduction

The dimensionality reduction techniques usually involve of both search algorithms and scoring functions.

Search algorithms generate subsets solutions possible, and they compare them by using the score as a measure

of the effectiveness of each solution. But the search for optimal solution is unattainable due to calculation cost.

Aha [9] cites three typical components to achieve a dimensionality reduction of features: a search algorithm to

traverse the space of subsets possible, an evaluation function to maximize, receiving a subset of features and

provides a numeric value; and finally a performance function which is here the classification.

D. Categorization of Dimensionality Reduction Methods According to the Feature Generation Process

Seen from the generation attributes, dimensionality reduction methods are either:

 They realize the reduction by transformation of data vectors which gives the attribute extraction,

Elkebir Sarhrouni et.al / International Journal of Engineering and Technology (IJET)

ISSN : 0975-4024

Vol 5 No 3 Jun-Jul 2013

2679

文档加载中……请稍候！
如果长时间未打开，您也可以点击刷新试试。

下载文档到电脑，查找使用更方便

10 玖币 0人已下载

立即下载

摘要：

ADashboardtoAnalysisandSynthesisofDimensionalityReductionMethodsinRemoteSensingElkebirSarhrouni#1,AhmedHammouch*2,DrissAboutajdine#3#LRIT,FacultyofSciences,MohamedV-AgdalUniversity,Morocco,A.IbnBattouta.4.B.P.:1014Rabat,Morocco1sarhrouni436@yahoo.fr*LRGE,ENSET,MohamedV-SouissiUniversity,MoroccoA.FAR...

展开>> 收起<<

A Dashboard to Analysis and Synthesis of Dimensionality Reduction Methods in Remote Sensing.pdf

共7页,预览2页

还剩页未读，继续阅读

声明：本站为文档C2C交易模式，即用户上传的文档直接被用户下载，本站只是中间服务平台，本站所有文档下载所得的收益归上传人(含作者)所有。玖贝云文库仅提供信息存储空间，仅对用户上传内容的表现方式做保护处理，对上载内容本身不做任何修改或编辑。若文档所含内容侵犯了您的版权或隐私，请立即通知玖贝云文库，我们立即给予删除！

A Dashboard to Analysis and Synthesis of Dimensionality Reduction Methods in Remote Sensing

相关推荐

开通VIP享超值会员特权

作者详情

相关内容

热门标签

举报选择: